CN111461159A

CN111461159A - Decoupling representation learning algorithm based on similarity constraint

Info

Publication number: CN111461159A
Application number: CN201910598166.0A
Authority: CN
Inventors: 李晓强; 陈亮波
Original assignee: University of Shanghai for Science and Technology
Current assignee: University of Shanghai for Science and Technology
Priority date: 2019-07-04
Filing date: 2019-07-04
Publication date: 2020-07-28

Abstract

The invention discloses a decoupling representation learning algorithm based on similarity constraint, and aims to solve the problem that the existing InfoGAN is in an unsupervised representation learning model. The method comprises the following specific steps: step one, preparing a data set; step two, selecting a model structure: in order to generate a picture with better visual effect, a WGAN structure with a gradient penalty is adopted; step three, applying similarity constraint to the model: constraints are imposed on the similarity and factors that generate the patches based on the WGAN structure using gradient penalties. Compared with an InfoGAN model, the similarity-constrained generation countermeasure network (SCGAN) is simple in structure, only the similarity constraint needs to be added on the basis of the original generation countermeasure network, meanwhile, the SCGAN has higher model robustness and still stably represents a processed data set, and the SCGAN is an unsupervised representation learning model, so that expensive labeling work can be avoided, and the application prospect is wide.

Description

Decoupling representation learning algorithm based on similarity constraint

Technical Field

The invention relates to the field of picture representation, in particular to a decoupling representation learning algorithm based on similarity constraint.

Background

In the probability statistics theory, the generation model is a model that can estimate a probability distribution from training data and randomly generate new observation data using the probability distribution. In order to make an excellent machine learning algorithm try to understand the intrinsic rules of the data, the algorithm needs to learn to create, that is, the generated data has important significance. The expression learning is a popular field in the generation learning, and is receiving attention of the learner. By representing the efficient representation obtained by learning, many discrimination tasks in machine learning, such as classification, segmentation, detection, etc., can be assisted. The decoupled representation learning belongs to a sub-branch of the representation learning, with the aim of learning factors that can control the high level semantic information of the picture. For supervised models, such as CGAN (conditional generation countermeasure network), labels are explicitly provided for factors that learn to control the class of items. For unsupervised models, such as InfoGAN (information maximization generation countermeasure network), the relationship between factors and picture representation is measured through mutual information, the lower bound of the mutual information is maximized by using a variation technology, and the factors are further learned to control the potential representation of the picture, such as illumination, color and the like.

Conditional generation models require the provision of tags for representation learning, in most cases, obtaining tags is expensive, and because conditional generation models provide tags for representations to be learned, the representations captured by the models are limited, such as capturing the number types on a handwritten character set.

InfoGAN is relatively classical in the unsupervised representation learning model. The idea of this model is to maximize the mutual information between the factor and the picture, and the intuitive explanation is that since the factor can control some representation of the picture, the factor must have a close relation to the picture, and the mutual information can be used to measure the relation. However, the model of the InfoGAN is complex, and in order to maximize mutual information, the InfoGAN uses a variational technique, and a neural network is additionally added to maximize the lower bound. Moreover, the InfoGAN training is not stable, and the collapse phenomenon is easy to occur on some processed data sets (randomly translated handwritten character data sets), so people are also researching in relevant aspects.

Disclosure of Invention

An object of the embodiments of the present invention is to provide a decoupling representation learning algorithm based on similarity constraint, so as to solve the problems proposed in the above background art.

In order to achieve the above purpose, the embodiments of the present invention provide the following technical solutions:

a decoupling representation learning algorithm based on similarity constraint specifically comprises the following steps:

step one, preparing a data set;

secondly, the training process of the WGAN is more stable than the original GAN and is easier to converge, in order to enable the discriminator to evaluate the Wassertein distance, the discriminator needs to be restricted under the limit of 1-L ipischitz (Lipschitz), and the essence of the constraint of 1-L ipitz is that the change degree of the penalty output of the discriminator is required to be smaller than the change degree of the input penalty, and some technologies can enable the discriminator to approximately reach the limit of 1-L ipitz, such as weight and gradient, in the actual operation, the effect of the gradient is far better than that of the penalty, and the clipping technology is preferably selected because the gradient is selected;

step three, applying similarity constraint to the model: on the basis of a WGAN structure using gradient punishment, the similarity and the factors of the generated pictures are constrained, the core idea is that when the factors are the same, the representation difference of the generated pictures should be as small as possible, otherwise, when the factors are different, the representation difference of the generated pictures should be as large as possible, the similarity constraint is that through the repulsion action generated among different factors, the attraction action is generated among the same factors, so that different factors can control different representations, and the purpose of representation learning is achieved.

As a further scheme of the embodiment of the invention: the data set in the step one comprises a simple data set and a complex data set, and can have rich samples, so that the most comprehensive samples are provided for the model to the maximum extent.

As a further scheme of the embodiment of the invention: the simple data set comprises MNIST, Fashinon-MNIST and SVHN, and the target is centered; the data set SVHN contains colored digital pictures, the background color is richer, and a plurality of numbers can appear in each picture.

As a further scheme of the embodiment of the invention: the complex data set comprises CIFAR-10 and CelebA, the data set CIFAR-10 contains pictures of real scenes, 10 types of objects are contained, and even if the same type of objects exist, huge differences can exist.

As a further scheme of the embodiment of the invention: since the difference in representation of the generated picture is not easily obtained, it is proposed to assume that the generated picture is composed of two parts of content and representation, and when the control content is the same, the difference in representation of the generated picture is equivalent to the difference in similarity of the generated picture, thereby converting the constraint from the difference in representation of the generated picture to the difference in similarity of the generated picture.

As a further aspect of the embodiments of the present invention. The two data sets MNIST and fast-MNIST contain a gray scale map, a black background and an object, which is centered.

As a further scheme of the embodiment of the invention: the data set CelebA contains a large number of star faces and has rich face representation.

Compared with the prior art, the embodiment of the invention has the beneficial effects that:

compared with an InfoGAN model, the SCGAN model is simple in structure, only the similarity constraint is added on the basis of an original generated countermeasure network, the SCGAN (the similarity constrained generated countermeasure network) has higher model robustness and still shows stability in a processed data set, and the SCGAN is also an unsupervised representation learning model, so that expensive labeling work can be avoided, and the application prospect is wide.

Drawings

Fig. 1 is a workflow diagram of embodiment 1 in a decoupling representation learning algorithm based on similarity constraints.

Detailed Description

The technical solution of the present patent will be described in further detail with reference to the following embodiments.

Example 1

step 1, noise z is sampled from Gaussian distribution at random, a factor c is sampled from polynomial distribution or uniform distribution, the factor c can be a discrete factor or a continuous factor, and the factor c is input into a generator to generate a picture.

And 2, randomly sampling a batch of real pictures x from the data set and sampling a batch of generated pictures G (z, c) from the output of the generator, and inputting the pictures into a discriminator to distinguish between true and false.

And 3, applying similarity constraint to the factor c and the generated picture G (z, c).

And 4, updating the parameters of the discriminator and the generator respectively by using a gradient descent algorithm.

Step 5, repeating steps 1-4 until the generator is able to generate a sufficiently realistic picture and different factors c are able to control different representations of the picture.

The working principle of the embodiment of the invention is as follows: the method is based on generation of a confrontation network and similarity constraint, and can learn the decoupling representation unsupervised. Because of the unsupervised learning mode, the pictures in the data set do not need to be labeled in advance. The representation of the picture is a relatively fuzzy concept, and it is very difficult for people to identify definite division boundaries, such as the division of the brightness of the picture. In an unsupervised mode, the potential characteristics of the data set can be automatically analyzed, and changes of more obvious picture representations can be learned.

The model structure of the method is very simple, and only the similarity constraint needs to be added on the basis of the original generation of the countermeasure network. And another unsupervised model InfoGAN for decoupling and representing learning needs to additionally introduce a neural network to optimize the lower bound of mutual information, so that the complexity of the model is increased. In addition, from the aspect of training difficulty, the model of the method is easier to converge, particularly on some processed data sets (handwriting character data sets which are translated randomly), the InfoGAN is broken down in the training process, and the model of the method is very robust.

The model of the method has better expandability. Since the similarity constraint only acts between the generated picture and the factor, any new type of generation countermeasure network model can integrate our similarity constraint. Therefore, the method can generate a picture with higher quality by using the novel models, and meanwhile, the similarity constraint can ensure that the factor captures the change expressed by the picture.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim concerned.

Furthermore, it should be understood that although the present description refers to embodiments, not every embodiment may contain only a single embodiment, and such description is for clarity only, and those skilled in the art should integrate the description, and the embodiments may be combined as appropriate to form other embodiments understood by those skilled in the art.

Claims

1. A decoupling representation learning algorithm based on similarity constraint is characterized by comprising the following specific steps:

step one, preparing a data set;

step two, selecting a model structure: structures employing WGAN using gradient penalties;

step three, applying similarity constraint to the model: constraints are imposed on the similarity and factors that generate the patches based on the WGAN structure using gradient penalties.

2. The similarity constraint-based decoupled representation learning algorithm of claim 1, wherein the first-step data set comprises a simple data set and a complex data set.

3. The similarity constraint-based decoupled representation learning algorithm of claim 2, wherein the simple dataset comprises MNIST, Fashion-MNIST, and SVHN.

4. The similarity constraint-based decoupled representation learning algorithm according to claim 2 or 3, characterized in that the complex dataset comprises CIFAR-10 and CelebA.

5. The similarity constraint-based decoupled representation learning algorithm of claim 1, wherein the generated picture is composed of two parts, namely content and representation.

6. The decoupled representation learning algorithm based on similarity constraints of claim 3 wherein the MNIST and Fashion-MNIST data sets contain gray scale maps, black backgrounds and objects.

7. The decoupling representation learning algorithm based on similarity constraint according to claim 4, wherein the data set CelebA contains a large number of star faces.